A simulation experiment of a pipeline based on machine learning for neutral hydrogen intensity mapping surveys

TL;DR

This paper demonstrates a machine learning-based pipeline for neutral hydrogen intensity mapping surveys, effectively handling various noise sources and RFI, and explores the thresholds for signal and interference amplitudes affecting performance.

Contribution

It introduces the first application of machine learning algorithms in HI intensity mapping pipelines, analyzing their effectiveness across different telescope configurations and interference conditions.

Findings

Pipeline performance improves significantly above certain sRFI amplitude thresholds.

Performance is largely independent of telescope aperture sizes.

Effective signal detection depends on specific amplitude thresholds.

Abstract

We present a simulation experiment of a pipeline based on machine learning algorithms for neutral hydrogen (HI) intensity mapping (IM) surveys with different telescopes. The simulation is conducted on HI signals, foreground emission, thermal noise from instruments, strong radio frequency interference (sRFI), and mild RFI (mRFI). We apply the Mini-Batch K-Means algorithm to identify sRFI, and Adam algorithm to remove foregrounds and mRFI. Results show that there exists a threshold of the sRFI amplitudes above which the performance of our pipeline enhances greatly. In removing foregrounds and mRFI, the performance of our pipeline is shown to have little dependence on the apertures of telescopes. In addition, the results show that there are thresholds of the signal amplitudes from which the performance of our pipeline begins to change rapidly. We consider all these thresholds as the edges of the signal amplitude ranges in which our pipeline can function well. Our work, for the first time, explores the feasibility of applying machine learning algorithms in the pipeline of IM surveys, especially for large surveys with the next-generation telescopes.